Python is slow and does not scale, right?
To get things running, you need
- docker
- the
makecommand, e.g. by installingapt install make
First, build the container with the service by running make build.
To run the benchmark, run make bench, which launches the benchmark.sh script.
Note that benchmark.sh launches multiple ab (Apache Benchmark) processes. This
happens because ab does not scale over multiple cores and would otherwise become
a bottleneck itself. You need to manually add up the results. On powerful machines
(~8 CPU cores) you need to launch a third ab process. Edit benchmark.sh to do
this (or to change any of the other benchmark parameters).
Since ab is runs on the same machine as the web server itself, part of the system's
resources are spent on on ab. However, this facilitates
| AWS instance type | concurrent connections | vCPUs | physical cores | requests / seconds |
|---|---|---|---|---|
| c5.large | 1250 | 2 | 1 | 4000 |
| c5.xlarge | 2500 | 4 | 2 | 8300 |
| c5.2xlarge | 5000 | 8 | 4 | 18000 |
| c5.4xlarge | 10000 | 16 | 8 | 34500 |
Note that on c5.4xlarge, I had to start three ab processes.
gevent, a coroutine-based concurrency library for Python